The study is currently in progress.
Proposed by: Prof. Markus Seidel
Immune dysregulation, including autoimmunity, autoinflammation, allergy, and malignancy predisposition, adds significant disease burden in primary immune disorders (PID) and inborn errors of immunity (IEI). Recently, a pilot version of this study evaluated whether the 5-graded immune deficiency
and dysregulation activity (IDDA2.1) score profile of 21 organ involvement and disease burden parameters could aid in diagnosing a broad spectrum of IEI. Supervised machine learning methods (support vector machine, logistic regression, random forest) were used for classification, ranking features by predictive power. Unsupervised clustering (k-nearest neighbors) and uniform manifold approximation and projection (UMAP) were applied to cluster and visualize disease-specific profiles. Genetic disorder occurrence was predicted with 72% balanced accuracy, CVIDs without proven gene defect at 76%, specific genetic disorders at 43%, IEI categories at 35%, and “cardinal genes” at 63%. A best-of-three approach improved genetic predictions to 70% accuracy. UMAP relatively clearly distinguished various IEI based solely on IDDA2.1 scores entered into the ESID registry by documentarists in 84 centers across Europe. Limited sample sizes for rare diseases emphasized the need for global collaboration and active data
contribution.
A new, larger, more comprehensive and ambitious study now
aims to integrate not only a newly implemented IDDA2.2 phenotype profile module, but importantly, the entire level 1 dataset, which was moderately amended in the ESID registry version 4.0 as of late 2024. Using ESID registry data as training datasets, we are planning to apply several statistical approaches including artificial neural networks (ANNs) and the
recent developments of its variations such as transformer, residual, diffusion and graph neural networks. The aim of the study is to extract the hidden features and their non-linear associations with respect to the disease profiles and phenotypes. Ultimately, endeavors like this should allow using patient registry datasets to develop electronic monitoring and
management guidance assistants.
